A real-time high-precision visual-inertial SLAM algorithm in dynamic environments

Visual simultaneous localization and mapping technology has achieved promising results in static scenes, yet it still encounters significant challenges in dynamic scenes. A prevalent approach in dynamic SLAM is to detect potential dynamic objects in the environment and eliminate dynamic features using deep learning. Unfortunately, deep learning algorithms relying on pixel-level object segmentation pose challenges for real-time operation. While methods combining deep-learning based object detection with multi-view geometry can achieve real-time performance, those based on multi-view geometry often presuppose precise camera poses. Given the difficulty in balancing accuracy and real-time performance of existing methods in dynamic environments, this paper introduces a high-precision and real-time visual-inertial SLAM algorithm. Our proposed algorithm employs multiple threads running in parallel, including object detection and dynamic object classification, feature tracking, local state optimization. Within the object detection and dynamic object classification thread, detected objects are categorized based on their relationship with prior dynamic objects, and further segmentation of dynamic objects within this class is performed using depth information. We propose an optical flow tracking method to robustly track dynamic objects, addressing situations of missed detections. Experiments conducted on the TUM RGB-D public dataset demonstrate that, with GPU acceleration, our proposed algorithm achieves a balance between accuracy and real-time performance in dynamic environments.